Machine Learning Applied to a Dual-Polarized Sentinel-1 Image for Wind Retrieval of Tropical Cyclones

In this work, three types of machine learning algorithms are applied for synthetic aperture radar (SAR) wind retrieval in tropical cyclones (TCs), and the optimal method is confirmed. total, 30 Sentinel-1 (S-1) images dual-polarization (vertical–vertical [VV] vertical–horizontal [VH] were collected during period from 2016 to 2021, which acquired interferometric-wide extra-wide modes with pixels 10 m 40 m, respectively. More than 100,000 sub-scenes a spatial coverage 3 km extracted these images. The dependences variables estimated sub-scenes, i.e., VV-polarized VH-polarized normalized cross-section (NRCS), as well azimuthal wave cutoff wavelength, on speeds stepped-frequency microwave radiometer (SFMR) soil moisture active passive (SMAP) studied, showing linear relations between speed parameters; however, saturation NRCS wavelength observed. This foundation selecting input algorithms. Two-thirds collocated dataset (20 images) used training process using algorithms, eXtreme Gradient Boosting (XGBoost), Multi-layer Perceptron, K-Nearest Neighbor, coefficients fitted after completion through 20 SFMR SMAP data. Another taken validation up 70 m/s, yielding 2.53 m/s root mean square error (RMSE) 0.96 correlation 0.12 scatter index (SI) XGBoost. result better >5 achieved existing cross-polarized geophysical model function other two algorithms; moreover, comparison retrievals XGBoost Level-2 CyclObs products shows about 4 RMSE 0.18 SI. suggests that algorithm an effective inverting TC field utilizing SAR measurements dual-polarization.